A branch-and-bound algorithm for convex multi-objective Mixed Integer Non-Linear Programming Problems Valentina Cacchiani 1 Claudia D Ambrosio 2 1 University of Bologna, Italy 2 École Polytechnique, France Recent advances in multi-objective optimization, Wien 2014 Acknowledgments to COST Action TD1207
Table of contents 1 Convex multi-objective MINLPs 2 Branch-and-bound 3 Preliminary computational experiments 4 Conclusions and Future research
Convex multi-objective MINLPs Convex multi-objective MINLPs min f k (x) k {1,..., p} (1) g i (x) 0 i {1,..., m} (2) x j Z j {1,..., r} (3) n is the number of variables r is the number of general integer variables (r n) f k, g i : R n R are twice continuously differentiable (non-linear) convex functions
Convex multi-objective MINLPs Literature review Branch-and-bound algorithms for multi-objective (bi-objective) MILPs Mavrotas and Diakoulaki 1998 Mavrotas and Diakoulaki 2005 Belotti, Soylu and Wiecek 2013 Vincent, Seipp, Ruzika, Przybylski and Gandibleux 2013 Parragh and Tricoire 2014 Stidsen, Andersed and Dammann 2014 Multi-objective (bi-objective) NLPs Fernández and Tóth 2007 Leyffer 2009 Ehrgott, Shao and Schöbel 2011
Branch-and-bound Branch-and-bound algorithm branching scheme dual bounds fathoming rules refinement procedure
Branch-and-bound Branch-and-bound algorithm Branching scheme: At each level j of the decision tree, we generate one child node for each possible fixing of variable x j to value l, with l {ub j,..., lb j } Dual bounds: The lower bound at the root node is computed by solving p single objective MINLP problems via a general-purpose MINLP solver. At each node of the decision tree, the lower bound is computed by solving p single objective NLP problems obtained by relaxing integrality requirements and by taking into account the branching decisions up to the current node.
Branch-and-bound Fathoming rules A node can be fathomed if: the corresponding relaxed problem is infeasible it is an integer feasible leaf node its lower bound is dominated by (at least) one of the solutions, say x, of the current Pareto set, i.e., LB k f k (x ) k {1,..., p} each single objective NLP k problem (k {1,..., p}) is integer feasible and all the p integer solutions coincide
Branch-and-bound Starting Pareto set and solving leaf nodes Weighted Sum method: min p λ k f k (x) k=1 g i (x) 0 i {1,..., m} x j Z j {1,..., r}, with 0 λ k 1 k {1,..., p} and p k=1 λ k = 1. Since we consider convex problems, the solution of the leaf nodes can generate all Pareto points by varying the weights (Censor 1977).
Branch-and-bound Refinement procedure For each solution x in the current Pareto set Y and for each objective function f k ( k {1,..., p}), we solve the following model with f k set to f k (x ). min f k (x) g i (x) 0 i {1,..., m} f k (x) f k x j Z k {1,..., p}, k k j {1,..., r}.
Preliminary computational experiments Preliminary Computational experiments: Hydro Unit Commitment & Scheduling A unit commitment problem of a generation company (Borghetti, D Ambrosio, Lodi, Martello 2008): find the optimal scheduling (maximize the power selling revenue) of a multiunit pump-storage hydro power station, for a short term period in which the electricity prices are forecast during the time horizon, a set of units can be: used as turbines to produce power used as pumps to pump water in the reservoir switched off several physical and operational constraints are imposed lower and upper bounds on the flows in the turbines limits on the flow variations in two consecutive periods water spillage to startup a pump or a turbine in Borghetti et al. lower bound on the final reservoir volume
Preliminary computational experiments Hydro Unit Commitment & Scheduling: MINLP model Binary variables are used to model the units behavior Continuous variables model the water flow passing through turbines or pumped by pumps and the water volume in the reservoir Additional variables are used to model the physical and operational constraints Bi-objective model: maximization of the revenue obtained from power selling: non-linear concave function maximization of the reservoir volume at the end of the time horizon: linear function
Preliminary computational experiments A discontinuous Pareto set Consider a single period of the time horizon and fix each of the 3 configurations (turbine on, pump on, both off): the Pareto set is the union of the three disjoint sets. 1500 1000 500 Revenue 0-500 -1000-1500 2.095e+07 2.1e+07 2.105e+07 2.11e+07 2.115e+07 Final reservoir volume
Preliminary computational experiments Characteristics of the instances # T: number of time periods of one hour considered in the instance # T # vars # bin # constr 1 18 8 19 2 30 14 34 3 42 20 49 4 54 26 64 5 66 32 79 6 78 38 94 7 90 44 109
Preliminary computational experiments Computational experiments: setting AMPL environment Intel Xeon 2.4 GHz with 8 GB Ram running Linux SCIP to solve single objective MINLPs Ipopt to solve single objective NLPs Weighted Sum method to obtain a starting Pareto set (step 0.1) Weighted Sum method to solve a leaf node (step 0.1)
Preliminary computational experiments Comparison Comparison of three branch-and-bound versions: norf: no refinement 1RF: refinement procedure only executed at the end of the resolution RF: refinement procedure executed at each update of the Pareto set
Preliminary computational experiments Comparison of three branch-and-bound versions Number of solutions CPU time (sec) # T norf 1RF RF norf 1RF RF 1 4 4 4 1 1 1 2 11 11 11 3 3 3 3 35 35 30 12 12 15 4 61 61 49 43 43 57 5 108 108 79 150 150 229 6 179 179 120 534 534 891 7 257 257 134 1946 1948 3861
Preliminary computational experiments Pareto sets of the three branch-and-bound versions 3000 2000 Start norf 1RF RF 1000 Revenue 0-1000 -2000-3000 2.06e+07 2.07e+07 2.08e+07 2.09e+07 2.1e+07 2.11e+07 2.12e+07 2.13e+07 2.14e+07 Final reservoir volume
Preliminary computational experiments Fathoming statistics # T # nodes # dom # leaf 1 12 1 1 2 55 1 5 3 233 4 19 4 862 11 65 5 3056 26 211 6 10415 54 665 7 34185 175 1995
Preliminary computational experiments Comparison with the Weighted Sum method (T=3) The Weighted Sum method: was executed with a step of 0.001, i.e. executed for 1000 iterations ended up in obtaining 27 solutions solutions are characterized by a high revenue and a limited final reservoir The branch-and-bound algorithm derives a more diverse Pareto set. The RF solutions are characterized by solutions having revenue and volume in wider ranges.
Preliminary computational experiments Comparison with the Weighted Sum method (T=3) 3000 2000 WS norf 1RF RF 1000 Revenue 0-1000 -2000-3000 2.06e+07 2.07e+07 2.08e+07 2.09e+07 2.1e+07 2.11e+07 2.12e+07 2.13e+07 2.14e+07 Final reservoir volume
Conclusions and Future research Conclusions We have presented a branch-and-bound algorithm for convex multi-objective MINLPs Preliminary computational experiments on instances of Hydro Unit Commitment & Scheduling show that the method finds a more diverse Pareto set compared to Weighted Sum method Future research will be devoted to compare the proposed method to the ɛ-constraint method improve the way of solving the leaf nodes and the fathoming rules to speed up the overall solution process test additional instances (e.g. convex nonlinear Knapsack Problem)
Conclusions and Future research Conclusions We have presented a branch-and-bound algorithm for convex multi-objective MINLPs Preliminary computational experiments on instances of Hydro Unit Commitment & Scheduling show that the method finds a more diverse Pareto set compared to Weighted Sum method Future research will be devoted to compare the proposed method to the ɛ-constraint method improve the way of solving the leaf nodes and the fathoming rules to speed up the overall solution process test additional instances (e.g. convex nonlinear Knapsack Problem) Thank you for your attention!